An AI-driven curtain wall modular design automatic generation method and system

Abstract

The application discloses an AI-driven curtain wall modular design automatic generation method and system, which comprises the following steps: acquiring real-time state data of a main structure to construct and update a built state digital twin model; extracting instantaneous boundary conditions of a to-be-installed module from the model; inputting the boundary conditions into an adaptive design generation model to generate adaptive adjustment parameters; applying the parameters to correct a standardized digital model to generate an adaptive module digital model that accurately matches the built state structure, and converting the adaptive module digital model into manufacturing instructions for output. Through automatic real-time design adjustment, the application solves the curtain wall installation conflict problem caused by dynamic deformation of the main structure, realizes accurate matching of the prefabricated component and the actual structure interface, improves construction efficiency and building quality, and reduces construction cost.

Description

Technical Field

[0001] This application relates to the field of computer-aided architectural design (CAD) technology, specifically to an AI-driven method and system for automatically generating modular curtain wall designs. Background Technology

[0002] In modern architectural practice, unitized curtain walls are widely used due to their high level of completion, high precision, and excellent water and air tightness. Their design and manufacturing process heavily relies on computer-aided architectural design and computer-aided manufacturing technologies. This involves high-precision prefabrication in a factory based on a fixed building information model, followed by transportation to the site for hoisting. In this scenario, the core of traditional computer-aided design (CAD) processes is the generation of design and manufacturing instructions based on a static, idealized digital model.

[0003] However, when faced with complex steel grid shell structures that are large-span, lightweight, and have free-form surfaces, such as the dome structures of large public transportation hubs or stadiums, existing technologies reveal insurmountable technical flaws. During construction, the geometry of such grid shell structures is not static. Due to the combined effects of various physical factors, including the structure's self-weight, temporary construction loads, welding stress release, and thermal expansion and contraction caused by changes in ambient temperature, partially completed structures continuously experience dynamic elastic deformations that are difficult to predict precisely using existing simulation technologies. This means that the three-dimensional spatial coordinates and normal vector orientation of each installation interface reserved for curtain wall units actually change in real time as construction progresses.

[0004] The dynamic uncertainty of the main structure and the static prefabrication production method of unitized curtain walls, which relies on static idealized models, cannot be well addressed with existing technologies. Curtain wall units, prefabricated based on ideal design models and with fixed dimensions, often exhibit severe geometric misalignment, torsion, and dimensional gaps at their interfaces with the already deformed actual structure during on-site installation, making successful installation impossible. Conventional solutions in the engineering field include allowing excessively large installation tolerances, using complex universal joints, and supplementing this with extensive on-site manual measurement, cutting, grinding, and secondary welding. This not only severely sacrifices the final accuracy and performance of the building but also leads to uncontrolled construction costs and schedules. Therefore, there is an urgent need in this field for a novel technical solution that can resolve the installation conflicts between dynamically changing structures and static prefabricated components. Summary of the Invention

[0005] This application provides an advanced computer-aided design (CAD) method and system that overcomes the limitations of existing CAD / CAM processes in handling dynamically changing structures by introducing artificial intelligence and digital twin technology.

[0006] To achieve the above objectives, one aspect of the present invention provides an AI-driven method for automatically generating modular curtain wall designs, comprising the following steps:

[0007] Step 1: Construct and update the built-up digital twin model in real time. Using 3D laser scanning equipment, photogrammetry equipment, or fixed sensor networks deployed at the construction site, periodically or continuously acquire 3D point cloud data of the main structure and the position and orientation data of key nodes. Input the acquired data in real time into a pre-set digital twin model to update the model's geometric information, thereby obtaining a built-up digital twin model that accurately represents the actual spatial state of the main structure at the current moment. Here, "built-up state" refers to the actual physical form of the structure after experiencing various loads.

[0008] Step two: Extract the instantaneous boundary conditions of the module to be installed. In the built-up digital twin model, according to the preset curtain wall installation sequence, locate the target installation area of ​​one or more modules to be installed. Accurately extract the geometric information of all installation interfaces from this area. The geometric information includes the three-dimensional spatial coordinates, normal vectors, and edge contours of adjacent installed modules at each interface connection point. This set of geometric information is then determined as the instantaneous boundary conditions of the module to be installed.

[0009] Step 3: Generate adaptive adjustment parameters for the module to be installed. The instantaneous boundary conditions are used as input to a pre-trained adaptive design generation model. The adaptive design generation model performs inference calculations based on the instantaneous boundary conditions and outputs a set of adaptive adjustment parameters for the standard module design. Specifically, the adaptive design generation model is a deep neural network model trained using a reinforcement learning algorithm. Its training objective function maximizes the geometric fit between the adjusted module and the instantaneous boundary conditions, while minimizing the module's material usage and deformation.

[0010] Step four: Generate manufacturing data for the adaptive module. Retrieve the standardized digital model of the module to be installed and apply the adaptive adjustment parameters to perform geometric corrections. Preferably, the geometric corrections include non-uniform scaling of the overall module dimensions, adjusting the length and angle of the module's borders to match instantaneous boundary conditions, and translating or rotating the preset position and orientation of its back connectors, thereby generating an adaptive module digital model that differs from the standard module but precisely matches the built-in structure.

[0011] Step 5: Output manufacturing data and implement closed-loop control. The adaptive module digital model is converted into computer-aided manufacturing instructions that can be recognized by automated processing equipment and then output. After one module or a batch of modules is installed, steps 1 through 5 are repeated to form a closed-loop control process from on-site status perception and cloud-based intelligent design to flexible manufacturing. This allows the design of all subsequent modules to dynamically adapt to the cumulative deformation of the main structure.

[0012] Furthermore, to continuously optimize the decision-making ability and adaptability of the adaptive design generation model to specific projects, the method of this invention also includes an online model optimization step. Specifically, after one or a batch of modules are installed according to step five, the actual fitting accuracy data of the interface between the module and the main structure is collected using the three-dimensional laser scanning device or embedded sensors deployed on the installed modules. The actual fitting accuracy data may include the final joint width, the stress value of the connector, etc., and is used as a feedback signal for online incremental retraining or weight fine-tuning of the adaptive design generation model described in step three. For example, in a reinforcement learning model, results higher than a preset fitting accuracy can be used as a positive reward signal, and results lower than a preset fitting accuracy can be used as a penalty signal to adjust the model's policy network.

[0013] Accordingly, another aspect of the present invention provides an AI-driven automatic generation system for modular curtain wall design, comprising:

[0014] The data acquisition module is used to acquire real-time status data of the building's main structure;

[0015] A digital twin module, connected to the data acquisition module, is used to construct and update a digital twin model based on the real-time status data;

[0016] The design generation module is connected to the digital twin module and has an adaptive design generation model deployed within it. This model is used to generate adaptive adjustment parameters for the module to be installed based on the instantaneous boundary conditions extracted from the built-up digital twin model, and to generate an adaptive module digital model. Optionally, the design generation module is also used to receive the actual fitting accuracy data of the installed module and to optimize the adaptive design generation model online based on the data.

[0017] And a data output module, connected to the design generation module, for converting the adaptive module digital model into computer-aided manufacturing instructions and outputting them.

[0018] In summary, this application provides a closed-loop, data-driven adaptive design and manufacturing method and system. The core of this solution lies in dynamically updating the digital twin model using real-time collected site data, ensuring it accurately reflects the completed state of the building structure. Based on this completed state model, and through artificial intelligence model reasoning, a unique set of adaptive design parameters is generated for each module to be installed, precisely matching its instantaneous installation boundaries. Finally, this parameterized adaptive design is transformed into precise manufacturing instructions. In this way, this application transforms the traditional static, batch-based prefabrication process into a dynamic, personalized, and on-demand responsive flexible manufacturing process, thereby solving the problem of discrepancy between the ideal design model and construction reality.

[0019] The technical solution provided in this application has the following beneficial effects:

[0020] 1. It effectively solves the installation conflict problem of unitized curtain walls caused by dynamic deformation during the construction of complex structures such as large spans and free-form surfaces, and can achieve precise matching between prefabricated components and the actual structure interface.

[0021] 2. Automated real-time design adjustments replace a large amount of on-site manual measurement, cutting, and correction work in traditional processes, improving installation accuracy and construction efficiency, shortening project time, and reducing costs caused by rework and material waste.

[0022] 3. It improves the quality of the final building and ensures key performance indicators such as the airtightness and watertightness of the curtain wall system, making it possible to achieve cutting-edge architectural designs that were previously difficult to realize due to limitations in construction precision.

[0023] 4. By introducing an online learning mechanism, the system can self-optimize within a project cycle, continuously improve design accuracy, and show greater adaptability to the unique deformation patterns of specific projects. Attached Figure Description

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0025] Figure 1 This is a flowchart of an AI-driven automatic generation method for modular curtain wall design provided in an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram of the structure of an AI-driven automatic generation system for modular curtain wall design provided in an embodiment of the present invention.

[0027] Figure 3 This is a schematic diagram illustrating the relationship between the ideal design model, the built-up structure, and the adaptive module in an embodiment of the present invention.

[0028] Figure 4This is a schematic diagram illustrating the working principle of the adaptive design generation model in an embodiment of the present invention.

[0029] Figure 5 This is a schematic diagram of the geometric transformation of the adaptive adjustment parameters of the curtain wall module in an embodiment of the present invention.

[0030] Figure 6 This is a schematic diagram of the three-dimensional view of the completed steel structure reticulated shell and the instantaneous boundary condition calibration in an embodiment of the present invention. Detailed Implementation

[0031] The technical solution of the present invention will now be clearly and completely described in conjunction with the accompanying drawings and specific embodiments.

[0032] Example 1: This example provides an AI-driven method for automatically generating modular curtain wall designs, addressing the technical challenge of mismatch between prefabricated curtain wall modules and actual installation interfaces caused by dynamic deformation of the main structure during construction. (Refer to...) Figure 1 The method includes the following steps:

[0033] Step S1: Construct and update the built-up digital twin model in real time. Using 3D laser scanning equipment, photogrammetry equipment, or fixed sensor networks deployed at the construction site, periodically or continuously acquire 3D point cloud data of the main structure and the position and orientation data of key nodes. Input the acquired data in real time into a preset digital twin model to update the model's geometric information, thereby obtaining a built-up digital twin model that accurately represents the actual spatial state of the main structure at the current moment. Here, "built-up state" refers to the actual physical form of the structure after experiencing various loads.

[0034] Specifically, step S1 can be broken down into the following sub-steps: First, data acquisition is performed. For example, one or more high-precision 3D laser scanners are used to scan the currently completed main steel structure grid shell to obtain the original 3D point cloud data covering the target area. The scanning cycle can be set according to the construction progress and environmental changes, for example, after the hoisting of a large structural unit, or at the two times of day with the greatest temperature difference (such as early morning and afternoon). Second, data preprocessing is performed. The point cloud data obtained from multiple stations are stitched and registered using a common target sphere or feature points, and noise points are filtered out using algorithms such as Statistical Outlier Removal. Third, model updating is performed. The preprocessed point cloud data is aligned with the initial Building Information Model (BIM), and the Iterative Closest Point (ICP) algorithm or its variants are used to calculate the deviation of the actual structure from the theoretical model. Then, based on this deviation, the coordinates of the grid vertices or key nodes in the digital twin model are precisely adjusted to generate the built-up digital twin model. This step ensures that the digital model is no longer a static design blueprint, but a real-time, high-fidelity mapping of the physical entity in digital space, providing accurate geometric basis for subsequent adaptive design.

[0035] Alternatively, as an alternative, data acquisition can also be achieved through a network of displacement monitoring sensors or fiber Bragg grating (FBG) strain sensors pre-embedded at key structural nodes (such as member intersections). These sensors can continuously output high-frequency position and strain data, allowing structural analysis algorithms to infer the overall morphology of the structure and update the digital twin model in real time.

[0036] Step S2: Extract the instantaneous boundary conditions of the module to be installed. In the built-up digital twin model, according to the preset curtain wall installation sequence, locate the target installation area of ​​one or more modules to be installed. Accurately extract the geometric information of all installation interfaces from this area. The geometric information includes the three-dimensional spatial coordinates, normal vectors, and edge contours of adjacent installed modules at each interface connection point. This set of geometric information is then determined as the instantaneous boundary conditions of the module to be installed.

[0037] In a specific application scenario, suppose the curtain wall installation sequence plans to install module MK-101 next. The system first locates the installation position of MK-101 in the built-up digital twin model. This position is typically defined by four preset connection points P1, P2, P3, and P4. The system automatically queries and extracts the precise 3D coordinates (x, y, z) of these four points in the current built-up model. Simultaneously, it extracts the normal vectors (nx, ny, nz) of the installation base plane containing these four points. If the adjacent module MK-100 has already been installed, the system also extracts the edge contour lines adjacent to MK-101. These discrete coordinates, vectors, and contour lines together constitute a structured dataset, namely the instantaneous boundary conditions of module MK-101. The core technical effect of this step is to transform the macroscopic, complex overall structural deformation into precise and unique geometric constraint inputs for a single module to be designed, thus clarifying and quantifying the design problem.

[0038] Step S3: Generate adaptive adjustment parameters for the module to be installed. The instantaneous boundary conditions are used as input to a pre-trained adaptive design generation model. The adaptive design generation model performs inference calculations based on the instantaneous boundary conditions and outputs a set of adaptive adjustment parameters for the standard module design. Specifically, the adaptive design generation model is a deep neural network model trained using a reinforcement learning algorithm. Its training objective function maximizes the geometric fit between the adjusted module and the instantaneous boundary conditions, while minimizing the module's material usage and deformation.

[0039] Furthermore, referring to Figure 4The adaptive design generative model can employ an actor-critic architecture from deep reinforcement learning. The actor network receives the instantaneous boundary conditions (state) and outputs a set of continuous adaptive adjustment parameters (actions). The critic network evaluates the expected long-term reward (Q-value) obtained by taking the action in the current state. This set of parameters can be a structured vector, such as [dx, dy, dz, rx, ry, rz, s_len, s_wid, c1_dx,c1_dy, ..., ci_dx, ci_dy, ..., cn_dx, cn_dy], where [dx, dy, dz] represent the translation of the module as a whole on the x, y, and z axes, [rx, ry, rz] represent the rotation around each axis, s_len and s_wid represent the scaling ratio of the length and width, respectively, and [ci_dx, ci_dy] represent the two-dimensional fine-tuning of the position of the i-th connector (i from 1 to n, where n is the total number of connectors) in the local coordinate system. The training objective (reward function) of the model can be designed as: R = a * F - b * M - c * D, where F is the geometric fit score (e.g., the reciprocal of the root mean square error between the coordinates of the adjusted module vertex and the target mounting point), M is the material usage (e.g., the area or volume of the adjusted module), D is the degree of deformation (e.g., the sum of squares of the rate of change of the diagonal length of the module quadrilateral, to penalize excessive shear deformation), and a, b, and c are weight coefficients. Through training in a large number of simulated deformation scenarios, the model learns how to make optimal adjustment decisions that balance economy and structural rationality while satisfying the primary objective of geometric fit. This method transforms the design process from parametric modeling following fixed rules into an optimization process where an AI agent makes dynamic decisions based on the real-time environment, with decision-making speed and the ability to handle complex constraints far exceeding that of humans.

[0040] As an alternative, a Generative Adversarial Network (GAN) can be used to build this model. In this model, the generator receives instantaneous boundary conditions and generates an adjusted geometric model of the modules; the discriminator then determines whether the model is realistic (from a perfectly fitted library of examples) or generated. Through adversarial training, the generator can eventually produce highly realistic modules that perfectly match the boundary conditions.

[0041] Step S4: Generate an adaptive module digital model. Retrieve the standardized digital model of the module to be installed and apply the adaptive adjustment parameters to perform geometric corrections. Preferably, the geometric corrections include non-uniform scaling of the overall module size, adjusting the length and angle of the module's border to match instantaneous boundary conditions, and translating or rotating the preset position and orientation of its back connectors, thereby generating an adaptive module digital model that differs from the standard module but precisely matches the built-in structure. (Refer to...) Figure 3 In an ideal design model, standard modules have regular geometric shapes, while the built-up structure has irregular installation interface contours due to deformation. The adaptive modules generated in this step can be precisely embedded into these irregular contours.

[0042] For example, a standardized curtain wall module might be created in parametric design software such as Grasshopper for Rhino or Dynamo for Revit. All its geometric features, such as frame length, glass dimensions, and connector positions, are controlled by a set of core parameters. The adaptive adjustment parameters [dx, dy, ...] output in step S3 are directly input into this parametric model as new values ​​for the corresponding parameters. The software engine automatically recalculates and generates the geometry based on the new parameters, resulting in an adaptive digital model of the module that precisely matches the instantaneous boundary conditions. This process is fully automated, avoiding errors that might be introduced by manual model modifications and ensuring rapid design response.

[0043] Step S5: Output manufacturing data and achieve closed-loop control. The adaptive module digital model is converted into computer-aided manufacturing (CAM) instructions that can be recognized by automated processing equipment and output. After a module or a batch of modules is installed, steps S1 to S5 are repeated to form a closed-loop control process from on-site status perception, cloud-based intelligent design to flexible manufacturing, so that the design of all subsequent modules can dynamically adapt to the cumulative deformation of the main structure.

[0044] Specifically, the generated adaptive modular digital model is exported as a general 3D model format (such as STEP or IGES) or a 2D drawing format (such as DXF) specific to a particular manufacturing process. For example, the cutting length and angle data of the curtain wall frame profiles are sent to a CNC cutting machine, and the precise contour data of the glass panels are sent to a CNC waterjet cutting machine. These instructions drive automated equipment in the factory for flexible manufacturing. Once this batch of modules is installed on-site, the system restarts step S1 to scan the structure containing the newly installed modules and obtain the latest built-up model. This closed-loop control mechanism enables the system not only to cope with initial structural deformation but also to compensate in real time for subsequent cumulative deformation caused by local stress redistribution or weight increase due to the installation of new modules, ensuring global consistency in the installation accuracy of the entire curtain wall system.

[0045] Furthermore, the method of the present invention also includes an online model optimization step. After one or a batch of modules are installed according to step S5, the actual fitting accuracy data of the interface between the module and the main structure is collected using the three-dimensional laser scanning device or an embedded sensor deployed on the installed module. The actual fitting accuracy data may include the final joint width, the stress value of the connector, etc., and the actual fitting accuracy data is used as a feedback signal for online incremental retraining or weight fine-tuning of the adaptive design generation model described in step S3. For example, in a reinforcement learning model, results higher than the preset fitting accuracy (e.g., joint width less than 2mm) can be used as positive reward signals, and results lower than the preset fitting accuracy (e.g., joint width greater than 5mm) can be used as penalty signals. The (state, action, actual reward) data is stored in the experience replay pool for fine-tuning of the model's policy network. The principle of this step is that it establishes a secondary feedback loop from the physical world to the digital model, enabling the AI ​​model to learn from the installation effect in the real world and continuously evolve itself. This allows the model to adapt to material properties or deformation patterns unique to a specific project that were not fully covered in the initial offline training, thereby achieving continuous improvement in design accuracy.

[0046] Example 2: This example provides an AI-driven automatic generation system for modular curtain wall design. (Refer to...) Figure 2 The system is a hardware and functional modular implementation of the method described in Example 1, including:

[0047] The data acquisition module 10 is used to acquire real-time status data of the main building structure. Physically, this module can consist of one or more 3D laser scanners, industrial camera arrays, or a sensor network composed of numerous GNSS receivers and strain gauges deployed on the construction site.

[0048] The digital twin module 20, connected to the data acquisition module 10, is used to construct and update the built-up digital twin model based on the real-time status data. This module is typically a software system running on a cloud server or a field edge computing node, integrating algorithms such as point cloud processing, model registration, and geometric updates. Its function corresponds to step S1 in Embodiment 1.

[0049] The design generation module 30, connected to the digital twin module 20, houses an adaptive design generation model. This model generates adaptive adjustment parameters for the module to be installed based on instantaneous boundary conditions extracted from the established digital twin model, and generates an adaptive digital model of the module. This module is the core intelligent hub of the system, running a pre-trained deep learning model (such as the aforementioned reinforcement learning model) and executing steps S2, S3, and S4 in Example 1. Optionally, the design generation module 30 also receives actual fitting accuracy data of the installed module and performs online optimization of the adaptive design generation model based on this data, thus performing online learning.

[0050] And a data output module 40, connected to the design generation module 30, is used to convert the adaptive module digital model into computer-aided manufacturing instructions and output them. This module is responsible for converting the design results into instructions that can be executed by manufacturing equipment, such as generating computer numerical control (CNC) code or DXF files, and transmitting them to the prefabricated component processing plant via the network, thus completing the function of step S5 in embodiment 1.

[0051] Those skilled in the art will understand that the connections between the modules of the above system can be electrical, fiber optic, or wireless communication connections. The system can be deployed on a cloud server to communicate with field devices and processing plants via a network, or it can be deployed as a distributed system that includes field edge computing units.

[0052] Example 3: This example uses the dome curtain wall project of a large stadium as an example to further illustrate the specific application process of the method of the present invention. The roof of the stadium adopts a double-layer free-form reticulated steel structure with a span of over 300 meters. Its geometry is extremely complex and requires extremely high construction precision. During construction, due to settlement caused by the self-weight of the steel structure, stress redistribution during the segmented hoisting process, and thermal expansion and contraction caused by temperature differences caused by sunlight, the actual spatial position of the reticulated shell nodes will have a cumulative deviation at the centimeter level compared with the initial design model, which is difficult to predict completely accurately using traditional finite element simulation software. If the traditional method is used, that is, all curtain wall units are prefabricated at once according to the initial design model, a lot of installation conflicts will inevitably occur with the interfaces of the already deformed actual structure after being transported to the site. Using conventional methods such as leaving too large tolerance joints or cutting and adjusting on site will seriously damage the overall air tightness, water tightness and final visual aesthetics of the building. Therefore, this project urgently needs a design and manufacturing method that can dynamically adapt to the actual shape of the structure.

[0053] In this embodiment, the curtain wall installation work of the project adopted the method described in this invention. First, after the hoisting and welding of the main steel structure in Zone C were completed, an automated total station deployed on site and a ground-based 3D laser scanner worked together to perform a high-precision scan of the interface area in Zone C where the first batch of curtain wall units would be installed, obtaining the 3D point cloud data of the area in its current built-up state. This data was transmitted in real time to a digital twin module deployed in the cloud via the on-site industrial network. The digital twin module compared the built-up point cloud with the original design model, calculating the deviation between the actual 3D coordinates and the theoretical coordinates of each installation base connection point. For example, it was found that the node numbered NC-101 was 18 mm lower than the design value in the Z-axis direction and 7 mm offset in the XY plane. Based on this, the system corrected the corresponding geometric information in the digital twin model, generating a built-up digital twin model that accurately reflects the current real state of the structure.

[0054] Next, according to the curtain wall installation plan, the system identifies the first unit to be installed, designated MC-001. The design generation module automatically extracts the precise three-dimensional coordinates of the four connecting nodes defining the installation position of MC-001, as well as the actual normal vector of the installation frame formed by these four points, from the updated built-up digital twin model. This set of data, containing precise coordinates and orientation, is formatted as an input vector, serving as the instantaneous boundary conditions for MC-001.

[0055] The instantaneous boundary condition vector is input into a pre-trained, deep reinforcement learning-based adaptive design generative model. This model completes inference within milliseconds, outputting a set of adaptively adjusted parameters for the MC-001 standard parameterized model. These parameters might specifically include: {boundary L1 length: -15.5mm; bounding box L2 length: -17.2mm; diagonal angle correction: +0.2 degrees; back connector A position offset: (x-3mm, y+2mm, z-18mm); ...}. These parameters are applied to the MC-001 standardized digital model, automatically generating an adaptive modular digital model with a non-standard quadrilateral outline and precisely adjusted back connector positions through parameterization.

[0056] Subsequently, the data output module automatically decomposes this adaptive module digital model, generating CAM (Computer-Aided Manufacturing) data to guide manufacturing. For example, it generates cutting instructions with precise lengths and end-cutting angles for the four aluminum alloy frames and sends them to the CNC cutting center; it generates precise irregular contour cutting paths for the glass panel and sends them to the waterjet cutting equipment. Based on this customized data, the processing plant produces only one unique unit, MC-001. After the unit arrives on site, it achieves millimeter-level precise alignment with nodes such as NC-101 and is successfully installed. After installation, on-site quality inspectors use a handheld 3D scanner to scan the installation joints, measuring the maximum gap width as 2.5 mm. This data is fed back to the design generation module as actual fitting accuracy data. Since 2.5 mm is slightly higher than the preset ideal threshold of 2 mm, this data (state, action, actual reward) is used to perform an online weight fine-tuning of the adaptive design generation model as a penalty signal, enabling it to generate a better adjustment strategy when dealing with similar deformation situations in the future. Subsequently, the system repeats the entire process described above to dynamically design and manufacture the next curtain wall unit MC-002 in Zone C, thus forming a closed-loop workflow of continuous perception, real-time design, precise manufacturing, and feedback optimization, ensuring the high-quality construction of the entire curved curtain wall.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An AI-driven method for automatically generating modular curtain wall designs, characterized in that, Includes the following steps: Step S1: Acquire the three-dimensional point cloud data of the main structure and the position and attitude data of key nodes through data acquisition equipment deployed at the construction site, and input the data into the preset digital twin model for updating, to obtain the completed digital twin model representing the real spatial state of the main structure at the current moment. Step S2: In the built-up digital twin model, according to the preset curtain wall installation sequence, locate the target installation area of ​​the module to be installed, and extract the geometric information of the installation interface from the area to determine the instantaneous boundary conditions of the module to be installed. Step S3: The instantaneous boundary conditions are used as input to a pre-trained adaptive design generation model. The adaptive design generation model performs inference calculations based on the instantaneous boundary conditions and outputs a set of adaptive adjustment parameters for a standard module. Step S4: Retrieve the standardized digital model of the standard module and apply the adaptive adjustment parameters to perform geometric correction on the standardized digital model to generate an adaptive module digital model that precisely matches the target installation area in the built-up digital twin model. The geometric correction includes non-uniform scaling of the overall size of the module, adjusting the length and angle of the module's border, and translating or rotating the preset position and orientation of its back connector. Step S5: Convert the adaptive module digital model into computer-aided manufacturing instructions and output them; after a module or a batch of modules is installed, repeat steps S1 to S5 to form a closed-loop control process.

2. The method according to claim 1, characterized in that, In step S1, the sub-step of inputting the data into the preset digital twin model for updating includes: aligning the three-dimensional point cloud data with the preset digital twin model, and using the iterative nearest point algorithm to calculate the deviation of the built-up digital twin model relative to the preset digital twin model, and adjusting the coordinates of the mesh vertices or key nodes in the preset digital twin model according to the deviation.

3. The method according to claim 1, characterized in that, In step S3, the adaptive design generation model is a deep neural network model trained based on a reinforcement learning algorithm. The objective function of its training process is used to maximize the geometric fit between the adjusted module and the instantaneous boundary conditions, while minimizing the amount of material used and the degree of deformation of the module.

4. The method according to claim 3, characterized in that, The deep neural network model employs an actor-critic architecture, where the actor network receives the instantaneous boundary conditions as the state and outputs the adaptive adjustment parameters as the action, and the critic network is used to evaluate the expected long-term reward that can be obtained by taking the action in the current state.

5. The method according to claim 1, characterized in that, It also includes an online model optimization step: after one or a batch of modules are installed according to step S5, the actual fitting accuracy data of the interface between the module and the main structure is collected, and the actual fitting accuracy data is used as a feedback signal to perform online incremental retraining or weight fine-tuning of the adaptive design generation model described in step S3.

6. The method according to claim 5, characterized in that, The actual fitting accuracy data includes at least one of the final joint width and the stress value of the connector; using the actual fitting accuracy data as a feedback signal for retraining or fine-tuning includes: using results with a higher than preset fitting accuracy as a positive reward signal and results with a lower than preset fitting accuracy as a penalty signal, so as to adjust the policy network of the adaptive design generation model.

7. The method according to claim 1, characterized in that, In step S3, the adaptive design generation model is a model built based on a generative adversarial network. The generator receives the instantaneous boundary conditions and generates an adjusted modular geometric model. The discriminator is used to determine whether the modular geometric model matches the instantaneous boundary conditions. Through adversarial training, the generator is able to produce a modular geometric model that matches the instantaneous boundary conditions.

8. The method according to claim 1, characterized in that, In step S1, the data acquisition device is a three-dimensional laser scanning device; after acquiring the three-dimensional point cloud data of the main structure, the step of preprocessing the three-dimensional point cloud data is also included. The preprocessing includes stitching and registering the point cloud data acquired from multiple stations, and filtering out noise points using a statistical outlier removal algorithm.

9. The method according to claim 1, characterized in that, In step S5, the computer-aided manufacturing instructions include the blanking length and cutting angle data of the curtain wall frame profile for the CNC cutting machine, and the contour data of the glass panel for the CNC waterjet cutting machine.

10. An AI-driven automatic generation system for modular curtain wall design, used to implement the method as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire 3D point cloud data of the main building structure and position and orientation data of key nodes; A digital twin module, connected to the data acquisition module, is used to construct and update a built-up digital twin model based on the data acquired by the data acquisition module. The built-up digital twin model can characterize the actual physical form of the main structure after experiencing load at the current moment. The design generation module, connected to the digital twin module, contains an adaptive design generation model. This model is used to: extract instantaneous boundary conditions of the module to be installed from the built-up digital twin model, including the three-dimensional spatial coordinates, normal vectors, and edge contours of adjacent installed modules at the installation interface; generate adaptive adjustment parameters for the module to be installed based on the instantaneous boundary conditions; retrieve a standardized digital model and apply the adaptive adjustment parameters to geometrically correct the standardized digital model, thereby generating an adaptive module digital model. And a data output module, connected to the design generation module, for converting the adaptive module digital model into computer-aided manufacturing instructions and outputting them.

Images